AI agent policy template

AI Agent Acceptable Use Policy: A Template for Companies

Your AI policy probably says what staff may type into a chatbot. It says nothing about software that signs in, clicks, sends email, buys and changes code on its own.

This template covers AI agents clause by clause, and pairs every clause with the control that enforces it.

AI Agent Acceptable Use Policy
1Purpose and scope
3Approved agents only
4Prohibited agents and features
7Human approval for actions
9MCP servers and tools
11Monitoring and enforcement
14 clauses, each with an enforcing control
Next step

AI Agent Blocklist: the enforcement layer for your policy

Block every agent your policy does not approve, on the network and on devices: 2,600+ agent domains and 5,300+ agent identifiers, updated every day.

From $99 a month for one organisation

Why a chatbot policy does not cover AI agents

An agent is a chatbot that was given hands. The questions a policy must answer change once software can act with an employee's access.

QuestionChat AI policyAI agent policy
What it controlsWhat people type and pasteWhat software does with company accounts
AccountsRarely mentionedWhich identities agents may use, and with which permissions
ActionsNone: the human actsWhich actions need a human to approve: pay, send, delete, deploy
ToolsNot relevantWhich extensions, MCP servers and plugins agents may load
Running timeOne question at a timeAgents that run for hours or on a schedule
AccountabilityThe person who used the answerThe person who started the agent, and the owner who approved it
EnforcementTraining and trustBlock lists, browser and device policies, permission rules

What goes wrong without an agent policy

These are reported events from 2025, described without names. Each one maps to a clause in the template below.

Clause 4

Hidden instructions

Researchers showed that hidden text in a web page comment could make an AI browser read a one-time code from the user's email and send it out.

Clause 9

Poisoned tools

A fake MCP server package copied every email sent through it to an outside address. It looked like a legitimate one.

Clause 8

Destructive actions

A coding agent deleted a production database during a declared code freeze.

Clause 7

No one approved it

Agents that send, buy or change records leave a gap: who decided, and who is responsible when it goes wrong?

Signs your current policy is out of date

  • It never uses the words agent, AI browser, extension or MCP.
  • It approves "AI tools" by brand, not by what they are allowed to do.
  • It says nothing about which accounts an AI tool may sign in with.
  • Developers add coding agents and MCP servers without a ticket.
  • Nobody can list the agents in use. See building an AI agent inventory.
  • No clause has a technical control behind it.

The template: 14 clauses for AI agents

Adapt the wording to your company and have legal and HR review it. Text in [brackets] is for you to fill in. Under each clause is the control that makes it real.

1

Purpose and scope

  1. This policy sets the rules for AI agents used for [Company] work, on any device or account.
  2. It applies to employees, contractors and anyone with access to company systems.
  3. It adds to, and does not replace, the general AI and IT acceptable use policies.
Control: HR acknowledgement at onboarding and each year.
2

Definitions

  1. AI agent: software that uses AI to take actions, not only to produce text.
  2. AI browser: a web browser with a built-in agent that can act on pages.
  3. Agent extension: a browser extension that reads or acts on pages using AI.
  4. MCP server or tool: a connector that gives an agent access to data or actions.
Control: the same definitions in your agent register.
3

Approved agents only

  1. Only agents listed in the approved agent register may be used for company work.
  2. Each approved agent has a named owner, an allowed purpose and allowed data classes.
  3. Approval of one product does not cover its new agent features; those need review.
Control: allowlist entries in browser, DNS and device policy.
4

Prohibited agents and features

  1. AI browsers may not be installed on company devices unless approved.
  2. Agent extensions are blocked unless on the approved list.
  3. Built-in browser agent features that act on pages are off unless approved.
Control: app blocking, extension block lists, Chrome GeminiActOnWebSettings.
5

Accounts and permissions

  1. Agents may not use administrator or privileged sessions.
  2. Staff may not give an agent their password or a personal access token without approval.
  3. Where possible, agents use their own identity with the least access needed.
Control: identity provider app consent rules; separate admin browsers.
6

Data rules

  1. Agents may not read or process [restricted data classes].
  2. Agents may not be pointed at systems holding customer personal data unless approved for that purpose.
  3. Agent output stays inside company systems.
Control: data classification labels and blocked agent domains.
7

Human approval for actions

  1. A person must approve before an agent pays, signs, sends external email, deletes data or changes access.
  2. Auto-approve settings that skip this step are not allowed.
  3. The person who approves is accountable for the action.
Control: approval settings in each approved agent, checked at review.
8

Coding and DevOps agents

  1. Coding agents run only on approved machines or sandboxes, never with production credentials.
  2. Agents may not read secrets files or credential stores.
  3. Shell commands that change systems need human approval.
Control: managed deny rules in the agent, for example a Claude Code managed-settings.json.
9

MCP servers and tools

  1. Agents may only use MCP servers and plugins from the approved list.
  2. Approved tools are pinned to a reviewed version.
  3. Public hubs and registries are not a source of approved tools.
Control: an MCP allowlist and blocked hub domains.
10

Buying and adding agents

  1. New agents, including agent features in existing software, go through [procurement or security review].
  2. The review covers data use, model training on company data, logging, data location and how to switch the agent off.
Control: the risk assessment checklist and the register.
11

Monitoring and enforcement

  1. [Company] blocks unapproved agents on its networks and devices and logs attempts.
  2. Logs are used to find unapproved use and improve the approved list, under [monitoring policy].
Control: DNS, firewall and device block lists, updated daily.
12

Incidents

  1. Report any agent action you did not expect, or any agent that asked for access it should not have, to [security contact] at once.
  2. Stop the agent and do not delete its history.
Control: incident runbook with an agent section.
13

Exceptions

  1. Exceptions are requested through [form], approved by [role], and have an end date.
  2. Exceptions are listed in the register with their scope.
Control: time-limited allow entries per group.
14

Review

  1. This policy is reviewed every [six] months and after any agent incident.
  2. Breaches are handled under [disciplinary policy].
Control: calendar owner and review log.

For how approvals, owners and reviews fit together across the company, see our guide to AI agent governance. For clause 5, see agent identity and permissions.

Agentic browsers and the same-origin policy: why clause 4 exists

Staff sometimes ask why an AI browser is riskier than a normal one with a chatbot tab. The answer is a browser rule they never see.

A normal browser

  • Code on one site cannot read your other sites. This is the same-origin policy.
  • A malicious page cannot open your email tab and read it.
  • Your logins are safe from other sites' code.

A browser with an agent

  • The agent is not site code. It acts as you, across every tab.
  • If a page tricks the agent with hidden text, the agent can open your email and act there.
  • The same-origin boundary does not stop it, because the agent is the user.

Gartner's December 2025 advice was titled "Cybersecurity Must Block AI Browsers for Now". The template bans AI browsers by default and lets you approve one through the exception process. More on the attack in AI browser prompt injection.

Policy as code: the control behind each clause

A clause that nothing enforces is a request. Write each rule down as a setting a system enforces, then the policy and the systems say the same thing.

ClauseControlWhere it lives
3 Approved onlyAllowlist of approved agents and extensionsBrowser policy, DNS allow rules, MDM app rules
4 AI browsersBlock install and launch of AI browser apps; block their service domainsIntune or Jamf app rules; DNS or firewall list
4 Agent extensionsExtension block list by IDChrome and Edge policy via Group Policy, Intune or Google Admin
4 Built-in agentsGeminiActOnWebSettings = 1, GeminiSettings as neededGoogle Admin console or Chrome policy
8 Coding agentsManaged allow, ask and deny rules; managed settings win over user filesThe agent's managed settings, deployed by MDM
8 Cloud coding agentsOutbound allowlist for the agent's environmentFor example the GitHub Copilot cloud agent firewall, under Settings, Copilot, Internet access
9 MCP toolsApproved MCP list; blocked MCP hub domainsAgent configuration; DNS or firewall list
11 MonitoringDaily block lists with loggingDNS, firewall, Defender indicators

Clause 9 in practice: building an MCP allowlist. Clause 8 in practice: stopping AI agents reading secrets. Clause 4 in practice: Chrome AI policies.

Agents inside business systems, privacy notices and regulation

Searches for an "AI agent policy" often mean one of three neighbouring things. The template covers the first and points to the other two.

Agents in your own platforms

ERP, procurement, claims and DevOps platforms now ship their own agents and "policy advisors". Apply the same clauses: a named owner, scoped permissions, approval thresholds for money and access, and review before new agent features go live.

An AI agent privacy policy

An acceptable use policy is an internal rule for staff. If your own agents process customers' personal data, you also need an external privacy notice that says so. Ask your privacy lead; it is a different document.

Frameworks and law

Align the policy with what your company already follows: the NIST AI Risk Management Framework, ISO/IEC 42001, the OWASP Top 10 for Agentic Applications (December 2025) and, in the EU, the AI Act. Map each clause to those you use.

Questions staff will ask, and short answers

Publish these with the policy. Most pushback comes from people who think the rule bans AI altogether. It does not: it bans unreviewed software acting with company access.

Staff questionAnswer to give
"Can I still use AI?"Yes. Approved assistants stay available. The policy covers agents that act on their own with your accounts.
"Why block my AI browser?"It acts with all your logins at once, and a web page can steer it with hidden text. Until that is solved, it stays off company devices.
"My extension only summarises pages."It still reads every page you open, including internal systems. Ask for it to be reviewed and added to the approved list.
"The coding agent saves me hours."Keep using an approved one, in a sandbox, with no production credentials and approval for commands that change systems.
"How do I get a tool approved?"Use the exception form. Reviews take [five working days], and approved tools go into the register.
"Is my use being watched?"Blocked attempts are logged to improve the approved list, as set out in the monitoring policy.

Most exposed are the people with the most access: finance, HR, support leads, developers and IT admins. Brief them first and in person, before the controls switch on.

Roll the policy out in six steps

A policy staff never see does nothing, and a control without a policy feels arbitrary. Do both together.

Step 1

Inventory

Find the agents already in use from DNS logs and device reports.

Step 2

Draft

Adapt the 14 clauses; fill in the bracketed parts.

Step 3

Review

Legal, HR, security and two business teams read it.

Step 4

Register

Publish the approved agent register with owners.

Step 5

Enforce

Switch on block lists and policies on the same day.

Step 6

Review

Check logs monthly and the policy every six months.

Before you add an agent to the register, run it through the AI agent risk assessment checklist. To choose between allowlist-first and blocklist-first enforcement, read allowlist or blocklist for AI agents.

Make clauses 3, 4, 9 and 11 enforce themselves

The AI Agent Blocklist turns "unapproved agents are not allowed" into block rules your firewall, DNS, browsers and MDM apply, and keeps them current every day.

  • 2,600+ AI agent domains as domain list, hosts, PAC, firewall and DNS filter files
  • Chrome and Edge extension block lists for Group Policy, Intune and Google Admin
  • macOS and Windows identifiers for Jamf and Intune
  • A category and risk grade on every entry, to match your register
  • Updated every day; access set up within one business day of an order

Ask us or order a plan

Tell us what you need to block and where your rules live. We reply within one business day.

  • Files for firewalls, DNS filters, secure web gateways, MDM and browser policies
  • Updated every day as new AI browsers, agents and extensions appear
  • Plans for one organisation, for MSPs and for product makers, or a hosted filter we run for you

We use your details only to answer you. No newsletters.

AI agent policy questions

What is an AI agent policy?

A written rule set for which agents staff may use, with which accounts and data, which actions need a human, and how the company enforces and reviews it. It covers AI browsers, agent extensions, coding agents, MCP servers and agents in business systems.

How is it different from an AI acceptable use policy?

A general AI policy covers what staff type into a chatbot. An agent policy adds what software may do on its own: sign in, click, send, buy, change code and call tools.

Should an AI agent security policy ban AI browsers?

Many companies ban them on company devices for now, in line with Gartner's December 2025 advice. The template bans them by default and approves specific tools through exceptions.

What does policy as code mean for AI agents?

Writing each clause as a setting a system enforces: browser policies, extension block lists, DNS and firewall lists, app control and permission rules inside coding agents.

Is an AI agent privacy policy the same thing?

No. The acceptable use policy is internal. A privacy notice tells customers how your own agents handle their personal data. You may need both.

How often should the policy be reviewed?

Every six months at least, and after any incident or major new agent type. Keep the register and block lists current between reviews; the lists change daily.

Claude and Claude Code are trademarks of Anthropic PBC. GitHub and Copilot are trademarks of GitHub, Inc. and Microsoft Corporation. Google, Chrome and Gemini are trademarks of Google LLC. Microsoft, Edge, Intune and Microsoft Defender are trademarks of Microsoft Corporation.

Jamf is a trademark of JAMF Software, LLC. Gartner is a trademark of Gartner, Inc. OWASP is a trademark of the OWASP Foundation. NIST is an agency of the U.S. Department of Commerce. ISO and IEC are the names of their standards organisations.

Names are used only to identify the products and frameworks discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these organisations.